Abstract Archives of the RSNA, 2010
Evaluation Methodology for an Automatic Multiple Sclerosis Lesion Detection and Quantification System in a Clinical Environment
Scientific Informal (Poster) Presentations
Presented on December 2, 2010
Presented as part of LL-INS-TH: Informatics
Kevin Chikai Ma BS, Presenter: Nothing to Disclose
James Reza F. Fernandez MD, MS, Abstract Co-Author: Nothing to Disclose
Jerry Tsu-Yuen Loo BS, Abstract Co-Author: Nothing to Disclose
Lilyana Amezcua MD, Abstract Co-Author: Nothing to Disclose
Alex Lerner MD, Abstract Co-Author: Nothing to Disclose
Mark S. Shiroishi MD, Abstract Co-Author: Nothing to Disclose
Brent Julius Liu PhD, Abstract Co-Author: Nothing to Disclose
Kathleen A. Garrison PhD, Abstract Co-Author: Nothing to Disclose
An automatic MS lesion quantification has been designed and is in the clinical validation process. A CAD workstation is set up in the clinical environment, and readers are prompted to manually segment MS lesions. The validation results show that the MS CAD is a viable option for lesion measurements and can be used in longitudinal lesion tracking and other research studies.
MRI is the clinical standard for diagnosing and tracking multiple sclerosis. MS lesions are manually segmented and volume is estimated by trained experts. The process is tedious and suffers intra- and inter-operator variability. We have developed an automatic computer-aided detection (CAD) and quantification tool of MS lesions based on k-nearest neighbors (KNN) algorithm, which assesses lesion probability on the voxel level. The method is able to successfully segment lesions in 3-D space. We have set up a validation process in the clinical environment to assess the CAD outputs.
KNN algorithm depends on a training set that must be created by manual segmentation of trained experts. Ten training cases are collected and lesions are manually segmented by two neuroradiologists. Six features are extracted: intensities of T1, T2, and FLAIR sequences, and 3-D spatial coordinates. Features of each voxel from reading cases are then inserted in the six-dimensional feature space to find k (k =100) nearest neighbors to determine the lesion voxel probability. To validate the algorithm, more than 100 ethnically diverse MS cases have been collected. A CAD workstation is implemented in USC Radiology. The workstation is equipped with the MS CAD program, a graphical user interface for case viewing and a manual segmentation tool. Trained experts would read an anonymized MS case and draw contours around lesions. The drawn contours are compared with CAD results. Time is recorded for both the manual and automated process to calculate efficiency.
Comparative results between CAD and manual lesion segmentation are similar and statistically significant, while the automatic segmentation process is more efficient. The validation study setup in the clinical environment allows access to clinical data within the PACS network.
Evaluation Methodology for an Automatic Multiple Sclerosis Lesion Detection and Quantification System in a Clinical Environment. Radiological Society of North America 2010 Scientific Assembly and Annual Meeting, November 28 - December 3, 2010 ,Chicago IL. http://archive.rsna.org/2010/9007545.html